As integrated circuits are produced with greater and greater levels of circuit density, efficient testing schemes that guarantee very high fault coverage while minimizing test costs and chip area overhead have become essential. However, as the complexity of circuits continues to increase, high fault coverage of several types of fault models becomes more difficult to achieve with traditional testing paradigms. This difficulty arises for several reasons. First, larger integrated circuits have a very high and still increasing logic-to-pin ratio that creates a test data transfer bottleneck at the chip pins. Second, larger circuits require a prohibitively large volume of test data that must be then stored in external testing equipment. Third, applying the test data to a large circuit requires an increasingly long test application time. And fourth, present external testing equipment is unable to test such larger circuits at their speed of operation.
Integrated circuits are presently tested using a number of structured design for testability (DFT) techniques. These techniques rest on the general concept of making all or some state variables (memory elements like flip-flops and latches) directly controllable and observable. If this can be arranged, a circuit can be treated, as far as testing of combinational faults is concerned, as a combinational network. The most-often used DFT methodology is based on scan chains. It assumes that during testing, all (or almost all) memory elements are connected into one or more shift registers, as shown in the U.S. Pat. No. 4,503,537. A circuit that has been designed for test has two modes of operation: a normal mode, and a test or scan mode. In the normal mode, the memory elements perform their regular functions. In the scan mode, the memory elements become scan cells that are connected to form a number of shift registers called scan chains. These scan chains are used to shift a set of test patterns into the circuit and to shift out circuit, or test, responses to the test patterns. The test responses are then compared to fault-free responses to determine if the circuit-under-test (CUT) works properly.
Scan design methodology has gained widespread adoption by virtue of its simple automatic test pattern generation (ATPG) and silicon debugging capabilities. Today, ATPG software tools are so efficient that it is possible to generate test sets (a collection of test patterns) that guarantee almost complete fault coverage of several types of fault models including stuck-at, transition, path delay faults, and bridging faults. Typically, when a particular potential fault in a circuit is targeted by an ATPG tool, only a small number of scan cells, e.g., 2-5%, must be specified to detect the particular fault (deterministically specified cells). The remaining scan cells in the scan chains are filled with random binary values (randomly specified cells). This way the pattern is fully specified, more likely to detect some additional faults, and can be stored on a tester.
Because of the random fill requirement, however, the test patterns are grossly over-specified. These large test patterns require extensive tester memory to store and a considerable time to apply from the tester to a circuit-under-test. FIG. 1 is a block diagram of a conventional system 18 for testing digital circuits with scan chains. External automatic testing equipment (ATE), or tester, 20 applies a set of fully specified test patterns 22 one by one to a CUT 24 in scan mode via scan chains 26 within the circuit. The circuit is then run in normal mode using the test pattern as input, and the test response to the test pattern is stored in the scan chains. With the circuit again in scan mode, the response is then routed to the tester 20, which compares the response with a fault-free reference response 28, also one by one. For large circuits, this approach becomes infeasible because of large test set sizes and long test application times. It has been reported that the volume of test data can exceed one kilobit per single logic gate in a large design. The significant limitation of this approach is that it requires an expensive, memory-intensive tester and a long test time to test a complex circuit.
These limitations of time and storage can be overcome to some extent by adopting a built-in self-test (BIST) framework, as shown in the U.S. Pat. No. 4,503,537. In BIST, additional on-chip circuitry is included to generate test patterns, evaluate test responses, and control the test. In conventional logic BIST, where pseudo-random patterns are used as test patterns, 95-96% coverage of stuck-at faults can be achieved provided that test points are employed to address random-pattern resistant faults. On average, one to two test points may be required for every 1000 gates. In BIST, all responses propagating to observable outputs and the signature register have to be known. Unknown values corrupt the signature and therefore must be bounded by additional test logic. Even though pseudo-random test patterns appear to cover a significant percentage of stuck-at faults, these patterns must be supplemented by deterministic patterns that target the remaining, random pattern resistant faults. Very often the tester memory required to store the supplemental patterns in BIST exceeds 50% of the memory required in the deterministic approach described above. Another limitation of BIST is that other types of faults, such as transition or path delay faults, are not handled efficiently by pseudo-random patterns. Because of the complexity of the circuits and the limitations inherent in BIST, it is extremely difficult, if not impossible, to provide a set of specified test patterns that fully covers hard-to-test faults.
Weighted pseudo-random testing is another method that is used to address the issue of the random pattern resistant faults. In principle, this approach expands the pseudo-random test pattern generators by biasing the probabilities of the input bits so that the tests needed for hard-to-test faults are more likely to occur. In general, however, a circuit may require a very large number of sets of weights, and, for each weight set, a number of random patterns have to be applied. Thus, although the volume of test data is usually reduced in comparison to fully specified deterministic test patterns, the resultant test application time increases. Moreover, weighted pseudo-random testing still leaves a fraction of the fault list left uncovered. Details of weighted random pattern test systems and related methods can be found in a number of references including U.S. Pat. Nos. 4,687,988; 4,801,870; 5,394,405; 5,414,716; and 5,612,963. Weighted random patterns have been primarily used as a solution to compress the test data on the tester. The generation hardware appears to be too complex to place it on the chip. Consequently, the voluminous test data is produced off-chip and must pass through relatively slow tester channels to the circuit-under-test. Effectively, the test application time can be much longer than that consumed by the conventional deterministic approach using ATPG patterns.
Several methods to compress test data before transferring it to the circuit-under-test have been suggested. They are based on the observation that the test cubes (i.e., the arrangement of test patterns bits as they are stored within the scan chains of a circuit-under-test) frequently feature a large number of unspecified (don't care) positions. One method, known as reseeding of linear feedback shift registers (LFSRs), was first proposed in B. Koenemann, "LFSR-Coded Test Patterns For Scan Designs," Proc. European Test Conference, pp. 237-242 (1991). Consider an .sub.n -bit LFSR with a fixed polynomial. Its output sequence is then completely determined by the initial seed. Thus, applying the feedback equations recursively provides a system of linear equations depending only on the seed variables. These equations can be associated with the successive positions of the LFSR output sequence. Consequently, a seed corresponding to the actual test pattern can be determined by solving the system of linear equations, where each equation represents one of the specified positions in the test pattern. Loading the resultant seed into the LFSR and subsequently clocking it will produce the desired test pattern. A disadvantage of this approach, however, is that seed, which encodes the contents of the test cube, is limited to approximately the size of the LFSR. If the test cube has more specified positions than the number of stages in LFSR, the test cube cannot be easily encoded with a seed. Another disadvantage of this approach is the time it requires. A tester cannot fill the LFSR with a seed concurrently with the LFSR generating a test pattern from the seed. Each of these steps must be done at mutually exclusive times. This makes the operation of the tester very inefficient, i.e., when the seed is serially loaded to the LFSR the scan chains do not operate; and when the loading of the scan chains takes place, the tester cannot transfer a seed to the LFSR.
Another compression method is based on reseeding of multiple polynomial LFSRs (MP-LFSRs) as proposed in S. Hellebrand et al., "Built-In Test For Circuits With Scan Based On Reseeding of Multiple Polynomial Linear Feedback Shift Registers," IEEE Trans. On Computers, vol. C-44, pp. 223-233 (1995). In this method, a concatenated group of test cubes is encoded with a number of bits specifying a seed and a polynomial identifier. The content of the MP-LFSR is loaded for each test group and has to be preserved during the decompression of each test cube within the group. The implementation of the decompressor involves adding extra memory elements to avoid overwriting the content of the MP-LFSR during the decompression of a group of test patterns. A similar technique has been also discussed in S. Hellebrand et al., "Pattern generation for a deterministic BIST scheme," Proc. ICCAD, pp. 88-94 (1995). Reseeding of MP-LFSRs was further enhanced by adopting the concept of variable-length seeds as described in J. Rajski et al., "Decompression of test data using variable-length seed LFSRs", Proc. VLSI Test Symposium, pp. 426-433 (1995) and in J. Raj ski et al., "Test Data Decompression for Multiple Scan Designs with Boundary Scan", IEEE Trans. on Computers, vol. C-47, pp. 1188-1200 (1998). This technique has a potential for significant improvement of test pattern encoding efficiency, even for test cubes with highly varying number of specified positions. The same documents propose decompression techniques for circuits with multiple scan chains and mechanisms to load seeds into the decompressor structure through the boundary-scan. Although this scheme significantly improves encoding capability, it still suffers from the two drawbacks noted above: seed-length limitations and mutually exclusive times for loading the seed and generating test patterns therefrom.
Thus, most reseeding methods to date suffer from the following limitations. First, the encoding capability of reseeding is limited by the length of the LFSR. In general, it is very difficult to encode a test cube that has more specified positions than the length of the LFSR. Second, the loading of the seed and test pattern generation therefrom are done in two separate, non-overlapping phases. This results in poor utilization of the tester time.
A different attempt to reduce test application time and test data volume is described in I. Hamzaoglu et al., "Reducing Test Application Time For Full Scan Embedded Cores," Proc. FTCS-29, pp. 260-267 (1999). This so-called parallel-serial full scan scheme divides the scan chain into multiple partitions and shifts in the same test pattern to each scan chain through a single scan input. Clearly, a given test pattern must not contain contradictory values on corresponding cells in different chains loaded through the same input. Although partially specified test cubes may allow such operations, the performance of this scheme strongly relies on the scan chain configuration, i.e., the number of the scan chains used and the assignment of the memory elements to the scan chains. In large circuits such a mapping is unlikely to assume any desired form, and thus the solution is not easily scalable. Furthermore, a tester using this scheme must be able to handle test patterns of different scan chain lengths, a feature not common to many testers.